Contextual Reranking of Search Engine Results
Abdullah Bas, Burak Gözütok, Kadriye Marangoz, Ersin Demirel, Hakki Yagiz Erdinc · 2022
Search engines are one of the most crucial components of social platforms to compete with rivals on the market. Nowadays, after the invention of BERT, contextual information became vital for almost every NLP problem. ElasticSearch is a successful third-party tool for platforms that do not have any search engine however, it is not sophisticated enough to respond to today’s needs. In this study, we aimed to enhance ElasticSearch by combining it with contextual information. We implemented this combination by adding a BERT-based Cross-Encoder (CE) on top of the ElasticSearch results for reranking to get better Mean Reciprocal Rank (MRR) scores. Taking into account the contextual information dramatically increased the existing ElasticSearch results and at the same time, it is pretty feasible at the deployment stage. Our method boosted ES results dramatically both in online and offline tests, especially on poor results of ElasticSearch. In the online A/B test, we enhanced our MRR scores by implementing this study by 77% in contrast to ElasticSearch and achieved a 0.23 MRR score.